Abstract
We present our approach to the AIMS-TBI 2026 Grand Challenge for automated traumatic brain injury (TBI) lesion segmentation and detection from T1-weighted MRI. Starting from baseline nnU-Net v2 models, we identified lesion contrast heterogeneity as the primary failure mode and developed a custom trainer with domain-specific contrast augmentations. Our final detection submission uses a 5-fold baseline ensemble with conservative post-processing, while segmentation uses a single-fold model fine-tuned with contrast augmentations on combined training and validation data (652 cases). Note that V2 differs from V1 in multiple factors (additional data, augmentations, fine-tuning), so individual contributions are not isolated. Our approach achieved a detection score of 0.8612 (Sensitivity 0.775, Specificity 0.947) and segmentation Dice of 0.4957. We analyze the progression from baseline to final submission, the impact of post-processing strategies, and lessons learned regarding the detection-segmentation trade-off.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AIMS_TBI_006.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/profile?id=~Sakshi_Rathi1
BibTex
@InProceedings{RatSak_DualModel_MICCAISAT2026,
author = { Rathi, Sakshi AND Sham, Yuk AND Samadani, Uzma},
title = { { Dual-Model nnU-Net Strategy for TBI Segmentation and Detection in T1-weighted MRI } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17254},
month = {pending},
page = {pending}
}
